Switch public uplift trees to the kernel backend, delete uplift.pyx (#955)

[jeongyoonlee]

What

Final issue of the tree-unification epic (#945): flip the public UpliftTreeClassifier / UpliftRandomForestClassifier to thin subclasses of the kernel-backed _KernelUpliftTreeClassifier / _KernelUpliftRandomForestClassifier (grown on the shared _tree Cython kernel), and delete the ~2.7 KLoC legacy uplift.pyx duplicate plus its import-time save/load monkey-patch. Closes #955.

Builds on the merged prereq #972 (native NaN support), #971 (#954 serialization + BaseEstimator conformance), and the rest of the epic (#946–#953).

Backward compatibility (names / args / defaults / predict semantics preserved)

• evaluationFunction still selects the criterion (KL/ED/Chi/CTS/DDP/IT/CIT/IDDP); mapped internally to the kernel criterion.
• Tree .predict(X) returns per-group P(Y=1|T=g) including control (shape (n, n_groups)), matching the historical return; absent-group leaf rates are zero-filled (uplift p_value is NAN #585).
• .fill(), .uplift_forest, .fitted_uplift_tree, and forest .predict(full_output=...) are all preserved.
• early_stopping_eval_diff_scale and fit's X_val / treatment_val / y_val are accepted for backward compatibility but ignored (the kernel has no validation-set early stopping) — a UserWarning is emitted when validation arrays are passed.
• DecisionTree stays importable via a DeprecationWarning shim.

Necessary behavior shims on the public subclasses

• feature_importances_ exposed non-negative + renormalized (the kernel's raw importances are signed, since an uplift split gain is not a monotone impurity decrease).
• int max_features clamped to n_features (legacy behavior; the kernel tree validates 0 < max_features <= n_features strictly) (shape mismatch when calling UpliftRandomForestClassifier.predict() #569).
• IDDP forces the honest approach at construction (legacy behavior).

Removed

• causalml/inference/tree/uplift.pyx (2742 lines) + its setup.py build entry + the tree/__init__.py monkey-patch.
• The legacy-vs-kernel parity unit tests and the uplift_tree_kernel_parity example notebook — obsolete now that the public classes are the kernel classes. tests/test_uplift_trees_kernel.py retains the kernel behavioral tests (criteria, regularization / normalization / honesty effects, pruning, plot, serialization, missing-value).

Verification

• tests/test_uplift_trees.py — 27 passed (acceptance; the one nodeSummary assertion migrated to the kernel's per-node group counts per the epic plan).
• tests/test_serialization_extended.py — passed.
• tests/test_uplift_trees_kernel.py — 48 passed (trimmed to kernel behavior).
• tests/test_causal_trees.py — passed (regression).
• black clean.

Notes

• The _uplift/ package was intentionally not renamed to uplift/ (the issue lists it as optional) to keep the diff surgical.
• Full sklearn check_estimator remains N/A (uplift fit is (X, treatment, y), not (X, y)), consistent with [Tree unification 8/9] Serialization + BaseEstimator + check_estimator; drop the monkey-patch #954's curated-conformance decision and documented in the class docstrings.

Part of #945.